Run a fleet of GPU or edge inference nodes and you learn the stack is brittle: one mismatch across kernel, NVIDIA driver, CUDA, firmware, runtime or, at the edge, JetPack turns a working node into a debugging session. Rebuilding that state through runtime configuration lets drift creep back, until every node is its own special case.

This talk argues for a different model: the machine becomes a versioned, reproducible image that can be shipped and rolled back atomically. When a vendor supports a specific Ubuntu-based stack, preserve it while adding an immutable lifecycle. Using Kairos as a concrete example, we will examine how to compose a reproducible AI node image, what belongs in it, and how this model changes upgrades, rollback and fleet visibility across datacenter and edge deployments.

If your AI stack needs machine-level compatibility, can your OS lifecycle reproduce it?